Community retests summarized by MaxForAI indicate that the performance, speed, and token efficiency improvements claimed by J-Space Cognition Suite for DeepSeek V4 currently lack reproducible supporting materials. The summary states that after GoForceX used an 87-question subset of Terminal-Bench 2.1 and loaded relevant skills and modules, the score actually decreased slightly, while Token Usage and Cost increased; the project primarily published aggregated numbers, lacking per-question logs, raw time consumption, token consumption, and a complete run manifest. As this is a compilation of project and community discussions by a regular account, it is currently more suitable as a signal requiring verification rather than a confirmed conclusion.<\/p>\n
Source:<\/p>\n
Available information shows that WeCom has begun offering CLI and MCP access to AI Agents, covering office capabilities such as messaging, email, documents, spreadsheets, tasks, calendars, meetings, WeDrive, and contacts, listing integration partners like Codex, DeepSeek Harness, Kimi Work, MiniMax Code, and WorkBuddy. MaxForAI compiled the open access page, and Gorden_Sun further referenced the WeCom announcement to discuss its CLI\/MCP upgrade. A noteworthy shift is that the competitive focus for Agents is moving from simply generating content to the ability to directly access real work data, contacts, and workflows; the specific available scope should still be confirmed via the official page.<\/p>\n
Sources:<\/p>\n
A Terminal-Bench 2.1 experiment shared by MaxForAI states that the single-run success rate for DeepSeek V4 Flash is approximately 79%. By having the model generate multiple candidate answers and then having the same model verify and select the best one, the result reached 88% \u00b1 0.6%, with costs reportedly 11 times lower than using a stronger judge model. Another compilation by MaxForAI notes that Qwen3.8-27B scored 52 points in a test, performing close to GLM5.2, which has about 28 times more parameters. Additionally, @oran_ge conducted a practical comparison of V4 Pro and Flash. Together, these point to a verifiable engineering direction: smaller models may not rely on single-pass capability to win, but can use lower inference costs to obtain more candidates and verification cycles; specific leaderboards and experimental setups should still be checked in the original repositories.<\/p>\n
Sources:<\/p>\n
The official Codex account explains that the Windows client now defaults to using Unified Exec, enabling it to start development servers, keep processes running, continuously observe output, send subsequent inputs, and run background test watchers, using the same command model as macOS\/Linux. Concurrently, Gorden_Sun referenced a preview of Claude Code’s `\/design` feature: the Agent can first generate multiple UI proposals, allowing the user to select a canvas, visually edit, and then implement. These two pieces of information, from a product release and user summary respectively, indicate that the scope of coding agents is expanding from one-time code generation to long-running tasks and bridging design with implementation.<\/p>\n
Sources:<\/p>\n
The Doubao Windows experience shared by dotey shows that users can view the execution progress of an Agent on their PC from their phone, temporarily interject with adjustments, and have it operate local documents, codebases, and graphical interfaces; the emphasized practical point is that PC resources remain on the PC, freeing the user from needing to stay by the device. Gorden_Sun shared usage patterns for Grok Bot controlling local Agents and cloud PCs, and using Feishu CLI to invoke Cursor\/Codex, while noting that cloud PC updates reset the system disk, causing interruptions to installed services. The former is a user’s practical test, the latter a user’s solution and its limitations, indicating that the value of remote Agents is emerging, but a stable, persistent environment remains a key bottleneck.<\/p>\n
Sources:<\/p>\n
Chengzilhy deconstructed an OOTD (Outfit of the Day) clothing change video: each outfit is generated as a separate 4-second vertical clip, uniformly using a “cup covers lens \u2014 cup removed to reveal \u2014 cup covers lens again” transition, then spliced together rhythmically to reduce the probability of losing control over character, scene, and clothing continuity when generating a long video in one go. Other tutorial-style content emphasizes first using a 3D pre-visualization stage to determine camera angles, actions, and shots before moving to Seedance for generation; Derek’s example also breaks down ad creation into controllable shots and prompts. These are bloggers’ practical tests and method sharing, with value in shifting from “repeated random draws” to a planned short-clip production pipeline.<\/p>\n
Sources:<\/p>\n
Derek_wall90176 proposed a method for outfit changes more suitable for e-commerce, advertising, and fashion proposals: the character reference image is responsible for identity, the garment image for structure, the wearing reference for drape, and the prompt for photography; it also requires preserving the collar, buttons, prints, fabric, cut, and body proportions, and suggests completing the basic composition first before adjusting the scene and lighting. 94vanAI, on the other hand, analyzed the boundaries of GPT-Image2 from the generation results, believing that illustration lighting and shadow are easier to make beautiful, while realistic human images, especially complex East Asian facial features and lighting, can still easily appear plastic. Both are bloggers’ actual tests and analyses, and their shared conclusion is: controllability is more important than the visual impact of a single image, and generated images still cannot replace real sizing and on-body judgment.<\/p>\n
Sources:<\/p>\n
NousResearch released the Bot Mode for Hermes Desktop: a Bot can have an independent role, model, memory, skills, and avatar, and multiple Bots can communicate with each other. The Pilot Harness shared by op7418 packages DeepSeek Harness into an out-of-the-box client and provides plugins for UI interaction, file tree, model service provider management, etc.; Ugur Aksay introduced Pascal Editor, an MIT-licensed open-source architecture\/3D design tool that runs in the browser and is advancing MCP to allow AI Agents to read and modify projects. The common change across different projects is that Agents are no longer just one-time chat interfaces but are organized into reusable work units with state, interfaces, and domain-specific tools.<\/p>\n
Sources:<\/p>\n
Statistics: Timeline Scanned Count=480 Matched Blogger Count=44 Matched Tweet Total=230 Weighted Tweet Score=189.8 Original Tweet Count=110 RT Tweet Count=34 Crawl Attempt Count=3 Boundary Coverage Status=tail_confidently_crossed_target_boundary<\/p>\n","protected":false},"excerpt":{"rendered":"
Today’s focus includes challenges in reproducing DeepSeek V4’s performance claims, WeChat Work opening office interfaces for AI Agents, and engineering paths for small models to improve efficiency through multiple generations and\u2026<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[19],"class_list":["post-1603","post","type-post","status-publish","format-standard","hentry","category-brief","tag-x--ai-"],"_links":{"self":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1603","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1603"}],"version-history":[{"count":0,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1603\/revisions"}],"wp:attachment":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1603"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1603"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1603"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}